U.S. patent application number 16/075540 was filed with the patent office on 2019-02-07 for offline identity authentication method and apparatus.
This patent application is currently assigned to GRG Banking Equipment Co., Ltd.. The applicant listed for this patent is GRG Banking Equipment Co., Ltd.. Invention is credited to Tiancai Liang, Dandan Xu, Yong Zhang.
Application Number | 20190042895 16/075540 |
Document ID | / |
Family ID | 57270226 |
Filed Date | 2019-02-07 |
![](/patent/app/20190042895/US20190042895A1-20190207-D00000.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00001.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00002.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00003.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00004.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00005.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00006.png)
![](/patent/app/20190042895/US20190042895A1-20190207-D00007.png)
![](/patent/app/20190042895/US20190042895A1-20190207-M00001.png)
![](/patent/app/20190042895/US20190042895A1-20190207-M00002.png)
![](/patent/app/20190042895/US20190042895A1-20190207-M00003.png)
View All Diagrams
United States Patent
Application |
20190042895 |
Kind Code |
A1 |
Liang; Tiancai ; et
al. |
February 7, 2019 |
OFFLINE IDENTITY AUTHENTICATION METHOD AND APPARATUS
Abstract
Provided are a method and a device for offline identity
authentication. The method includes: acquiring two or more images
for identity authentication to constitute a to-be-authenticated
multivariate image group; extracting a concatenated PCA convolution
feature of each of the images in the to-be-authenticated
multivariate image group, to obtain feature vectors; fusing
information on the images in the to-be-authenticated multivariate
image group based on the feature vectors according to a score
fusion strategy with a supervisory signal, to obtain a fusion
vector; and inputting the fusion vector to a pre-trained SVM
classifier to authenticate and determine whether the images in the
to-be-authenticated multivariate image group are consistent with
one another, to obtain an identity authentication result.
Inventors: |
Liang; Tiancai; (Guangzhou,
Guangdong, CN) ; Xu; Dandan; (Guangzhou, Guangdong,
CN) ; Zhang; Yong; (Changzhou, Guangdong,
CN) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
GRG Banking Equipment Co., Ltd. |
Guangzhou, Guangdong |
|
CN |
|
|
Assignee: |
GRG Banking Equipment Co.,
Ltd.
Guangzhou, Guangdong
CN
|
Family ID: |
57270226 |
Appl. No.: |
16/075540 |
Filed: |
June 12, 2017 |
PCT Filed: |
June 12, 2017 |
PCT NO: |
PCT/CN2017/087867 |
371 Date: |
August 3, 2018 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06K 9/00288 20130101;
G06N 3/0454 20130101; G06K 9/6288 20130101; G06K 9/46 20130101;
G06K 9/629 20130101; G06K 9/6273 20130101; G06N 3/08 20130101; G06K
9/4628 20130101; G06N 3/04 20130101; G06N 20/10 20190101; G06K
9/6247 20130101; G06K 9/6269 20130101; G06K 9/6215 20130101; G06K
9/64 20130101; H04L 63/0861 20130101 |
International
Class: |
G06K 9/64 20060101
G06K009/64; G06K 9/62 20060101 G06K009/62; G06K 9/00 20060101
G06K009/00; G06K 9/46 20060101 G06K009/46; H04L 29/06 20060101
H04L029/06; G06N 3/04 20060101 G06N003/04; G06N 3/08 20060101
G06N003/08 |
Foreign Application Data
Date |
Code |
Application Number |
Jun 12, 2016 |
CN |
201610410966.1 |
Claims
1. A method for offline identity authentication, comprising:
acquiring two or more images for identity authentication to
constitute a to-be-authenticated multivariate image group;
extracting a concatenated PCA convolution feature of each of the
images in the to-be-authenticated multivariate image group, to
obtain feature vectors; fusing information on the images in the
to-be-authenticated multivariate image group based on the feature
vectors according to a score fusion strategy with a supervisory
signal, to obtain a fusion vector; and inputting the fusion vector
to a pre-trained SVM classifier to authenticate and determine
whether the images in the to-be-authenticated multivariate image
group are consistent with one another, to obtain an identity
authentication result.
2. The method according to claim 1, wherein the multivariate image
group is a ternary image group, and comprises a chip image of an ID
card, a surface image of the ID card and a live face image.
3. The method according to claim 2, wherein a SVM classifier is
trained by following steps: acquiring a chip image of an ID card, a
surface image of the ID card and a live face image as samples to
constitute a ternary sample group in a training set, wherein the
ternary sample group in the training set comprises positive samples
and negative samples; extracting a concatenated PCA convolution
feature of each of the images in the ternary sample group, to
obtain sample feature vectors; fusing information on the images in
the ternary sample group based on the sample feature vectors
according to the score fusion strategy with the supervisory signal,
to obtain a sample fusion vector; and inputting the sample fusion
vector to the SVM classifier for training the SVM classifier, to
obtain the pre-trained SVM classifier.
4. The method according to claim 2, wherein the extracting the
concatenated PCA convolution feature of each of the images in the
to-be-authenticated ternary image group to obtain the feature
vectors comprises: inputting each of the images in the
to-be-authenticated ternary image group to a pre-trained deep
convolutional neural network; and selecting convolution outputs of
N intermediate sub-layers of convolution groups in the deep
convolutional neural network, as concatenated layers, and
sequentially performing PCA transform on the obtained concatenated
layers layer by layer to output the feature vector, wherein
N.gtoreq.2.
5. The method according to claim 4, wherein the deep convolutional
neural network comprises five convolution groups and two fully
connected layers, and each of the convolution groups comprises two
convolution sub-layers and one pooling layer, and wherein the
selecting convolution outputs of N intermediate sub-layers in the
deep convolutional neural network as concatenated layers and
sequentially performing PCA transform on the obtained concatenated
layers layer by layer to output the feature vector comprises:
extracting an output of the pooling layer of a fourth convolution
group, and stringing all values of the output into a first vector;
performing PCA transform on the first vector and reserving a first
number of principal components to obtain a first insertion vector;
extracting an output of the pooling layer of a fifth convolution
group, stringing all values of the output into a second vector, and
inserting the first insertion vector into a header of the second
vector; performing PCA transform on the inserted second vector, and
reserving a second number of principal components to obtain a
second insertion vector; extracting an output of a second fully
connected layer as a third vector, and inserting the second
insertion vector to a header of the third vector; and performing
PCA transform on the inserted third vector, and reserving a third
number of principal components to obtain the feature vector.
6. The method according to claim 2, wherein the fusing information
on the images in the to-be-authenticated ternary image group based
on the feature vectors according to the score fusion strategy with
the supervisory signal to obtain the fusion vector comprises:
calculating a cosine similarity degree between each pair of feature
vectors of the three feature vectors corresponding to the
to-be-authenticated ternary image group, as three matching scores;
calculating a difference between each of the matching scores and a
preset empirical threshold corresponding to the matching score, as
difference signals; encoding based on each of preset decision
weights and the difference signal corresponding to the preset
decision weight, to obtain weight signals, wherein the preset
decision weights have a one-to-one correspondence with three pairs
of the images in the to-be-authenticated ternary image group; and
synthesizing the matching scores, the difference signals and the
weight signals, as the fusion vector.
7. The method according to claim 6, wherein the encoding based on
each of the preset decision weights and the difference signal
corresponding to the preset decision weight to obtain the weight
signals comprises: converting a ratio of the preset decision
weights of three matching branches into an integer ratio, and
normalizing each of integers in the integer ratio to a range from 0
to 7, wherein the three matching branches comprise a matching
branch between the chip image and the surface image, a matching
branch between the chip image and the live face image, and a
matching branch between the surface image and the live face image;
converting the normalized integers in the integer ratio of the
decision weights of the matching branches into binary codes, to
obtain initial codes; and inserting a highest-order code
corresponding to each of the difference signals into the initial
code corresponding to the difference signal, to obtain the weight
signals, wherein the highest-order code corresponding to the
difference signal is one in a case where the difference signal is
greater than zero, and the highest-order code corresponding to the
difference signal is zero in a case where the difference signal is
less than or equal to zero.
8. A device for offline identity authentication, comprising: a
multivariate image acquiring module configured to acquire two or
more images for identity authentication to constitute a
to-be-authenticated multivariate image group; a convolution feature
extracting module configured to extract a concatenated PCA
convolution feature of each of the images in the
to-be-authenticated multivariate image group, to obtain feature
vectors; a score fusing module configured to fuse information on
the images in the to-be-authenticated multivariate image group
based on the feature vectors according to a score fusion strategy
with a supervisory signal, to obtain a fusion vector; and an
authenticating and determining module configured to input the
fusion vector into a pre-trained SVM classifier to authenticate and
determine whether the images in the to-be-authenticated
multivariate image group are consistent with one another, to obtain
an identity authentication result.
9. The device according to claim 8, wherein the multivariate image
group is a ternary image group, and comprises a chip image of an ID
card, a surface image of the ID card and a live face image.
10. The device according to claim 9, wherein the score fusing
module comprises: a matching score calculating unit configured to
calculate a cosine similarity degree of each pair of feature
vectors of the three feature vectors corresponding to the
to-be-authenticated ternary image group, as three matching scores;
a difference signal calculating unit configured to calculate a
difference between each of the matching scores and a preset
empirical threshold corresponding to the matching score, as
difference signals; a weight signal calculating unit configured to
encode based on each of preset decision weights and the difference
signal corresponding to the preset decision weight, to obtain
weight signals, wherein the preset decision weights have a
one-to-one correspondence with three pairs of the images in the
to-be-authenticated ternary image group; and a fusion vector
synthesizing unit configured to synthesize the matching scores, the
difference signals and the weight signals, as a fusion vector.
Description
[0001] This application claims priority to Chinese Patent
Application No. 201610410966.1, titled "OFFLINE IDENTITY
AUTHENTICATION METHOD AND APPARATUS", and filed with the Chinese
State Intellectual Property Office on Jun. 12, 2016, which is
incorporated herein by reference in its entirety.
FIELD
[0002] The present disclosure relates to the field of pattern
recognition, and in particular to a method and a device for offline
identity authentication.
BACKGROUND
[0003] When authenticating an identity based on an ID card, both
whether the ID card is genuine and whether a current user is a
legal holder of the ID card are identified. Whether the ID card is
genuine is determined using chip anti-counterfeiting technology
based on whether the ID card is machine-read successfully by a card
reader. The identity of the current user is checked and
discriminated online in combination with an ID card system of the
Ministry of Public Security, to further authenticate the identity.
On the one hand, the online authentication method is performed only
in a case where face database in the system of the Ministry of
Public Security is permitted to be accessed, which limits
application places. On the other hand, a fake ID card which has a
genuine chip and a fake surface image cannot be automatically
identified with the online authentication method. The fake ID card
may be machine-read successfully, and the identity is authenticated
to be genuine by online authentication. However, an image stored in
the chip is not consistent with the surface image of the ID card.
In this case, whether the chip image, the surface image of the ID
card and a holder image of the ID card are consistent with one
another is identified visually, to identify the fake ID card, which
undoubtedly increases burden on a checker and even results in false
detection or missed detection. In view of the problem, an
intelligent authentication method is required, in which, offline
authentication without relying on the face database of the Ministry
of Public Security can be implemented, and whether the chip image,
the surface image of the ID card and the holder image of the ID
card are consistent with one another is identified simultaneously,
to automatically provide an authentication result on whether the
authentication is passed, thereby improving authentication
efficiency.
SUMMARY
[0004] A method for offline identity authentication and a device
for offline identity authentication are provided according to the
embodiments of the present disclosure, to solve problems in the
conventional technology that the face database of the Ministry of
Public Security is relied on and it is difficult to identify
whether a chip image, a surface image of the ID card and a holder
image of the ID card are consistent with one another.
[0005] A method for offline identity authentication is provided
according to an embodiment of the present disclosure, which
includes: acquiring two or more images for identity authentication
to constitute a to-be-authenticated multivariate image group;
extracting a concatenated PCA convolution feature of each of the
images in the to-be-authenticated multivariate image group, to
obtain feature vectors; fusing information on the images in the
to-be-authenticated multivariate image group based on the feature
vectors according to a score fusion strategy with a supervisory
signal, to obtain a fusion vector; and inputting the fusion vector
to a pre-trained SVM classifier to authenticate and determine
whether the images in the to-be-authenticated multivariate image
group are consistent with one another, to obtain an identity
authentication result.
[0006] Optionally, the multivariate image group is a ternary image
group, and comprises a chip image of an ID card, a surface image of
the ID card and a live face image.
[0007] Optionally, a SVM classifier is trained by the following
steps: acquiring a chip image of an ID card, a surface image of the
ID card and a live face image as samples to constitute a ternary
sample group in a training set, where the ternary sample group in
the training set includes positive samples and negative samples;
extracting a concatenated PCA convolution feature of each of the
images in the ternary sample group, to obtain sample feature
vectors; fusing information on the images in the ternary sample
group based on the sample feature vectors according to the score
fusion strategy with the supervisory signal, to obtain a sample
fusion vector; and inputting the sample fusion vector to the SVM
classifier for training the SVM classifier, to obtain the
pre-trained SVM classifier.
[0008] Optionally, the extracting the concatenated PCA convolution
feature of each of the images in the to-be-authenticated ternary
image group to obtain the feature vectors includes: inputting each
of the images in the to-be-authenticated ternary image group into a
pre-trained deep convolutional neural network; and selecting
convolution outputs of N intermediate sub-layers of convolution
groups in the deep convolutional neural network as concatenated
layers, and sequentially performing PCA transform on the obtained
concatenated layers layer by layer to output the feature vector,
wherein N.gtoreq.2.
[0009] Optionally, the deep convolutional neural network includes
five convolution groups and two fully connected layers, and each of
the convolution groups includes two convolution sub-layers and one
pooling layer. The selecting convolution outputs of N intermediate
sub-layers in the deep convolutional neural network as concatenated
layers and sequentially performing PCA transform on the obtained
concatenated layers layer by layer to output the feature vector
includes: extracting an output of the pooling layer of a fourth
convolution group, and stringing all values of the output into a
first vector; performing PCA transform on the first vector and
reserving a first number of principal components to obtain a first
insertion vector; extracting an output of the pooling layer of a
fifth convolution group, stringing all values of the output into a
second vector, and inserting the first insertion vector into a
header of the second vector; performing PCA transform on the
inserted second vector, and reserving a second number of principal
components to obtain a second insertion vector; extracting an
output of a second fully connected layer as a third vector, and
inserting the second insertion vector to a header of the third
vector; and performing PCA transform on the inserted third vector,
and reserving a third number of principal components to obtain the
feature vector.
[0010] Optionally, the fusing information on the images in the
to-be-authenticated ternary image group based on the feature
vectors according to the score fusion strategy with the supervisory
signal to obtain the fusion vector includes: calculating a cosine
similarity degree between each pair of feature vectors of the three
feature vectors corresponding to the to-be-authenticated ternary
image group as three matching scores; calculating a difference
between each of the matching scores and a preset empirical
threshold corresponding to the matching score, as difference
signals; encoding based on each of preset decision weights and the
difference signal corresponding to the preset decision weight, to
obtain weight signals, wherein the preset decision weights have a
one-to-one correspondence with three pairs of the images in the
to-be-authenticated ternary image group; and synthesizing the
matching scores, the difference signals and the weight signals, as
the fusion vector.
[0011] Optionally, the encoding based on each of the preset
decision weights and the difference signal corresponding to the
preset decision weight to obtain the weight signals includes:
converting a ratio of the preset decision weights of three matching
branches into an integer ratio, and normalizing each of integers in
the integer ratio to a range from 0 to 7, wherein the three
matching branches includes a matching branch between the chip image
and the surface image, a matching branch between the chip image and
the live face image, and a matching branch between the surface
image and the live face image; converting the normalized integers
in the integer ratio of the decision weights of the matching
branches into binary codes, to obtain initial codes; and inserting
a highest-order code corresponding to each of the difference
signals into the initial code corresponding to the difference
signal, to obtain the weight signals, where the highest-order code
corresponding to the difference signal is one in a case where the
difference signal is greater than zero, and the highest-order code
corresponding to the difference signal is zero in a case where the
difference signal is less than or equal to zero.
[0012] A device for offline identity authentication is provided
according to an embodiment of the present disclosure, which
includes a multivariate image acquiring module, a convolution
feature extracting module, a score fusing module and an
authenticating and determining module. The multivariate image
acquiring module is configured to acquire two or more images for
identity authentication to constitute a to-be-authenticated
multivariate image group. The convolution feature extracting module
is configured to extract a concatenated PCA convolution feature of
each of the images in the to-be-authenticated multivariate image
group, to obtain feature vectors. The score fusing module is
configured to fuse information on the images in the
to-be-authenticated multivariate image group based on the feature
vectors according to a score fusion strategy with a supervisory
signal, to obtain a fusion vector. The authenticating and
determining module is configured to input the fusion vector into a
pre-trained SVM classifier to authenticate and determine whether
the images in the to-be-authenticated multivariate image group are
consistent with one another, to obtain an identity authentication
result.
[0013] Optionally, the multivariate image group is ternary image
group, and includes a chip image of an ID card, a surface image of
the ID card and a live face image.
[0014] Optionally, the score fusing module includes a matching
score calculating unit, a difference signal calculating unit, a
weight signal calculating unit and a fusion vector synthesizing
unit. The matching score calculating unit is configured to
calculate a cosine similarity degree of each pair of feature
vectors of the three feature vectors corresponding to the
to-be-authenticated ternary image group, as three matching scores.
The difference signal calculating unit is configured to calculate a
difference between each of the matching scores and a preset
empirical threshold corresponding to the matching score as
difference signals. The weight signal calculating unit is
configured to encode based on each of preset decision weights and
the difference signal corresponding to the preset decision weight,
to obtain weight signals, where the preset decision weights have a
one-to-one correspondence with three pairs of the images in the
to-be-authenticated ternary image group. The fusion vector
synthesizing unit is configured to synthesize the matching scores,
the difference signals and the weight signals, as a fusion
vector.
[0015] It may be seen from the above technical solutions that the
embodiments of the present disclosure have the following
advantages.
[0016] In the embodiments of the present disclosure, two or more
images for identity authentication are acquired to constitute a
to-be-authenticated multivariate image group. A concatenated PCA
convolution feature of each of the images in the
to-be-authenticated multivariate image group is extracted to obtain
feature vectors. Information on the images in the
to-be-authenticated multivariate image group is fused based on the
feature vectors according to a score fusion strategy with a
supervisory signal, to obtain a fusion vector. The fusion vector is
inputted into a pre-trained SVM classifier to authenticate and
determine whether the images in the to-be-authenticated
multivariate image group are consistent with one another, to obtain
an identity authentication result. In the embodiments of the
present disclosure, a chip image, a surface image of a certificate
such as an ID card and a live face image can be compared offline
from an overall decision, and an authentication result can be
provided based on a single decision of the pre-trained SVM
classifier, thereby reducing burden on the checker and improving
authentication efficiency without relying on face database of the
Ministry of Public Security.
BRIEF DESCRIPTION OF THE DRAWINGS
[0017] FIG. 1 is a flowchart of a method for offline identity
authentication according to an embodiment of the present
disclosure;
[0018] FIG. 2 is a flowchart of a method for offline identity
authentication according to another embodiment of the present
disclosure;
[0019] FIG. 3 is a schematic principle diagram of a method for
offline identity authentication according to an embodiment of the
present disclosure in an application scenario;
[0020] FIG. 4 is a schematic principle diagram of a deep
convolutional neural network model and concatenated PCA convolution
according to an embodiment of the present disclosure;
[0021] FIG. 5 is a schematic principle diagram of a score fusion
strategy with a supervisory signal according to an embodiment of
the present disclosure;
[0022] FIG. 6 is a schematic diagram showing encoding for a weight
signal according to an embodiment of the present disclosure;
[0023] FIG. 7 is a structural diagram of a device for offline
identity authentication according to an embodiment of the present
disclosure; and
[0024] FIG. 8 is a structural diagram of a device for offline
identity authentication according to another embodiment of the
present disclosure.
DETAILED DESCRIPTION OF EMBODIMENTS
[0025] The problems in the conventional technology that the face
database of the Ministry of Public Security is relied on and it is
difficult to identify whether the chip image, the surface image of
an ID card and a holder image are consistent with one another are
solved.
[0026] In order to make an objective, features and advantages of
the present disclosure clearer and easier to be understood, the
technical solutions according to the embodiments of the present
disclosure are described clearly and completely below in
conjunction with the drawings in the embodiments of the present
disclosure. It is apparent that the embodiments described below are
merely a part of the embodiments of the present disclosure, rather
than all the embodiments of the present disclosure. All other
embodiments obtained by those skilled in the art based on the
embodiments of the present disclosure without any creative work
fall within the protection scope of the present disclosure.
[0027] Referring to FIG. 1, a method for offline identity
authentication according to an embodiment of the present disclosure
includes steps 101 to 104.
[0028] In step 101, two or more images for identity authentication
are acquired to constitute a to-be-authenticated multivariate image
group.
[0029] The two images for identity authentication may include, for
example, a chip image of an ID card and a surface image of the ID
card, or may include the chip image of the ID card and a live face
image.
[0030] In step 102, a concatenated PCA convolution feature of each
of the images in the to-be-authenticated multivariate image group
is extracted to obtain feature vectors.
[0031] After the to-be-authenticated multivariate image group is
constituted, the concatenated principal component analysis (PCA)
convolution feature of each of the images in the
to-be-authenticated multivariate image group may be extracted, to
obtain the feature vectors.
[0032] In step 103, information on the images in the
to-be-authenticated multivariate image group is fused based on the
feature vectors according to a score fusion strategy with a
supervisory signal to obtain a fusion vector.
[0033] After the feature vectors are obtained, the information on
the images in the to-be-authenticated multivariate image group may
be fused based on the feature vectors according to the score fusion
strategy with the supervisory signal, to obtain the fusion
vector.
[0034] In step 104, the fusion vector is inputted into a
pre-trained SVM classifier to authenticate and determine whether
the images in the to-be-authenticated multivariate image group are
consistent with one another, to obtain an identity authentication
result.
[0035] After the fusion vector is obtained, the fusion vector may
be inputted to the pre-trained SVM classifier to authenticate and
determine whether the images in the to-be-authenticated
multivariate image group are consistent with another other, to
obtain the identity authentication result. For example, in a case
where the multivariate image group includes the chip image of the
ID card and the surface image of the ID card, whether the chip
image is consistent with the surface image is authenticated and
determined.
[0036] In the embodiment, two or more images for identity
authentication are acquired to constitute the to-be-authenticated
multivariate image group. A concatenated PCA convolution feature of
each of the images in the to-be-authenticated multivariate image
group is extracted to obtain feature vectors. Information on the
images in the to-be-authenticated multivariate image group is fused
based on the feature vectors according to the score fusion strategy
with the supervisory signal, to obtain a fusion vector. The fusion
vector is inputted into the pre-trained SVM classifier to
authenticate and determine whether the images in the
to-be-authenticated multivariate image group are consistent with
one another, to obtain an identity authentication result. In the
embodiment, the images for identity authentication can be compared
offline from an overall decision, and an authentication result may
be provided based on a single decision of the pre-trained SVM
classifier, thereby reducing burden on the checker and improving
the authentication efficiency without relying on the face database
of the Ministry of Public Security.
[0037] For ease of understanding, a method for offline identity
authentication according to an embodiment of the present disclosure
is described in detail below. Referring to FIG. 2, a method for
offline identity authentication (with taking authentication for an
image of an ID card as an example) according to another embodiment
of the present disclosure includes steps 201 to 213.
[0038] In step 201, a chip image of an ID card, a surface image of
the ID card and a live face image are acquired to constitute a
to-be-authenticated ternary image group.
[0039] The chip image of the ID card, the surface image of the ID
card and the live face image may be acquired to constitute a
to-be-authenticated ternary image group. The chip image may be
directly read by a card reading device. The surface image of the ID
card may be obtained by a scanning device. The live face image may
be obtained by a camera device. A face region in the acquired image
may be detected by a face detector.
[0040] It should be illustrated that, in the embodiment, three
images constitutes a multivariate image group, that is, a ternary
image group including the chip image of the ID card, the surface
image of the ID card and the live face image, and identity
authentication based on a chip of the ID card may be performed
using the ternary image group. In a case where the chip image, the
surface image and the live face image are determined to be
consistent with one another, the identity authentication is
passed.
[0041] In step 202, each of the images in the to-be-authenticated
ternary image group is inputted into a pre-trained deep
convolutional neural network.
[0042] After the to-be-authenticated ternary image group is
acquired, each of the images in the to-be-authenticated ternary
image group may be inputted into a pre-trained deep convolutional
neural network. As shown in FIG. 4, the deep convolutional neural
network includes several convolution groups and several fully
connected layers. Each of convolution groups includes several
convolution sub-layers and one pooling layer. In order to output a
reasonable feature vector, the deep convolutional neural network
includes five convolution groups and two fully connected layers in
the embodiment. When extracting the convolution feature, outputs
(pool4 and pool5 as shown in FIG. 4) of pooling layers of the last
two convolution groups of the five convolution groups and an output
(fc2 as shown in FIG. 4) of the last fully connected layer are
extracted, and PCA transform is performed on the outputs layer by
layer to reserve a principal component of the output of each
layer.
[0043] After the image is inputted into the deep convolutional
neural network, convolution outputs of N (N.gtoreq.2) intermediate
sub-layers are selected in the deep convolutional neural network as
concatenated layers, and PCA transform is performed on the obtained
concatenated layers sequentially layer by layer to output a feature
vector, which is described in detail with steps 203 to 208.
[0044] In step 203, an output of a pooling layer of a fourth
convolution group is extracted, and all values of the output are
strung into a first vector.
[0045] As shown in FIG. 4, after each of the images in the
to-be-authenticated ternary image group is inputted into a
pre-trained deep convolutional neural network, the output (pool4)
of the pooling layer of the fourth convolution group may be
extracted, and all values of the output are strung into a first
vector.
[0046] In step 204, PCA transform is performed on the first vector,
and a first number of principal components are reserved to obtain a
first insertion vector.
[0047] After all values of the outputs are strung into the first
vector, the PCA transform may be performed on the first vector, and
a first number of principal components are reserved to obtain the
first insertion vector (PP1). The first number may be set according
to a specific situation, for example, the first number of principal
components may be the first 1024 principal components.
[0048] In step 205, an output of a pooling layer of a fifth
convolution group is extracted, all values of the output are strung
into a second vector, and the first insertion vector is inserted
into a header of the second vector.
[0049] As shown in FIG. 4, after a first number of principal
components are reserved to obtain the first insertion vector, the
output (pool5) of the pooling layer of the fifth convolution group
may be extracted, all values of the output are strung into the
second vector, and the first insertion vector is inserted into the
header of the second vector.
[0050] In step 206, PCA transform is performed on the inserted
second vector, and a second number of principal components are
reserved to obtain a second insertion vector.
[0051] After the first insertion vector is inserted into the header
of the second vector, PCA transform may be performed on the
inserted second vector, and a second number of principal components
may be reserved to obtain the second insertion vector (PP2). The
second number may be set according to a specific situation, for
example, the second number of principal components may be the first
512 principal components.
[0052] In step 207, an output of a second fully connected layer is
extracted as a third vector, and the second insertion vector is
inserted into a header of the third vector.
[0053] After the PCA transform is performed on the inserted second
vector, and a second number of principal components are reserved to
obtain the second insertion vector, the output of the second fully
connected layer may be extracted as the third vector, and the
second insertion vector is inserted into the header of the third
vector (fc2).
[0054] In step 208, PCA transform is performed on the inserted
third vector, and a third number of principal components are
reserved to obtain a feature vector.
[0055] After the second insertion vector is inserted into the
header of the third vector, the PCA transform may be performed on
the inserted third vector, and a third number of principal
components may be reserved to obtain the feature vector (PP3). The
third number may be set according to a specific situation, for
example, the third number of principal components may be the first
256 principal components.
[0056] It should be understood that, for the three images in the
to-be-authenticated ternary image group, each of the images
corresponds to one feature vector, and the ternary image group
corresponds to three feature vectors.
[0057] In step 209, a cosine similarity degree between each pair of
feature vectors of the three feature vectors corresponding to the
to-be-authenticated ternary image group is calculated as three
matching scores.
[0058] After the feature vectors are obtained, the cosine
similarity degree between each pair of feature vectors of the three
feature vectors corresponding to the to-be-authenticated ternary
image group may be calculated as three matching scores. It should
be understood that one cosine similarity degree (the cosine
similarity degree may be used to evaluate a similarity degree
between two feature vectors) may be calculated between each pair of
feature vectors of the three feature vectors. Therefore, three
cosine similarity degrees may be calculated by pairwise combination
of the three feature vectors, and the cosine similarity degree is
used as a matching score between the two feature vectors
corresponding to the cosine similarity degree.
[0059] Specifically, the cosine similarity degree between two
images (I.sub.1, I.sub.2) is calculated according to the following
equation:
sim ( I 1 , I 2 ) = k = 1 n f 1 k f 2 k k = 1 n f 1 k 2 k = 1 n f 2
k 2 ##EQU00001##
[0060] where n denotes a dimension of a feature vector, f.sub.1k
denotes a k-th element of the feature vector of I.sub.1, and
f.sub.2k denotes a k-th element of the feature vector of I.sub.2.
According to the equation, a matching score between the chip image
and the surface image of the ID card is s.sub.1, and a matching
branch between the chip image and the surface image of the ID card
is denoted as a branch p1. A matching score between the chip image
and the live face image is s.sub.2, and a matching branch between
the chip image and the live face image is denoted as a branch p2. A
matching score between the surface image of the ID card and the
live face image is s.sub.3, and a matching branch between the
surface image of the ID card and the live face image is denoted as
a branch p3.
[0061] In step 210, a difference between each of the matching
scores and a preset empirical threshold corresponding to the
matching score is calculated as difference signals.
[0062] After the three matching scores are calculated, a difference
between each of matching score and a preset empirical threshold
corresponding to the matching score may be calculated as difference
signals.
[0063] It should be illustrated that the empirical threshold of
each of the matching branches in a training sample set (see details
in step 213) may be obtained using the 1:1 authentication
algorithm, and the empirical threshold is calculated according to
the following equation:
arg max T i = 1 m .delta. { ( s i - T ) .times. y i > 0 } i = 1
m i ##EQU00002##
[0064] where m denotes the number of sample pairs, s.sub.i denotes
a similarity degree of an i-th sample pair, and y.sub.i is a class
label of the i-th sample pair. In a case where the sample pair
corresponds to the same person, the class label of the sample pair
is 1. In a case where the sample pair corresponds to different
persons, the class label of the sample pair is -1. .delta.{}
denotes an indicator function and is defined as follows:
.delta. { x } = { 1 , x is true 0 , x is false . ##EQU00003##
[0065] The empirical thresholds T.sub.1, T.sub.2 and T.sub.3 of the
three matching branches are obtained.
[0066] In this step, the empirical threshold of each of the
matching branches is subtracted from the matching score of the
matching branch, to obtain difference signals of the three matching
branches, which may be d.sub.1=s.sub.1-T.sub.1,
d.sub.2=s.sub.2-T.sub.2 and d.sub.3=s.sub.3-T.sub.3.
[0067] In step 211, encoding is performed based on each of preset
decision weights and the difference signal corresponding to the
preset decision weight, to obtain a weight signal.
[0068] After the difference signals are obtained, encoding may be
performed based on each of the preset decision weights and the
difference signal corresponding to the preset decision weight, to
obtain the weight signals. The preset decision weights have a
one-to-one correspondence with three pairs of images in the
to-be-authenticated ternary image group. Specifically, step 211
includes the following steps (1) to (3).
[0069] In step (1), a ratio of the preset decision weights of the
three matching branches is converted into an integer ratio, and an
integer in the integer ratio is normalized to a range from 0 to 7,
The three matching branches includes a matching branch between the
chip image and the surface image of the ID card, a matching branch
between the chip image and the live face image, and a matching
branch between the surface image of the ID card and the live face
image.
[0070] In step (2), the normalized integer in the integer ratio of
the decision weights of the matching branches is converted into
binary codes, to obtain initial codes.
[0071] In step (3), a highest-order code corresponding to each of
the difference signals is inserted into the initial code
corresponding to the difference signal, to obtain weight
signals.
[0072] In a case where the difference signal is greater than zero,
the highest-order code corresponding to the difference signal is
one. In a case where the difference signal is less than or equal to
zero, the highest-order code corresponding to the difference signal
is zero.
[0073] For example, assuming that the ratio of the decision weights
of the branch p1, the branch p2 and the branch p3 is 5:3:2, the
ratio of the decision weights are encoded into 101, 011, and 010.
In a case where the difference signal of the current branch p1 is
positive, the highest-order bit of four bits for the branch p1 is
1, and the weight signal of the branch p1 is encoded into a binary
code 1101, and the binary code are converted into a decimal number
13. Therefore, the weight signals c.sub.1, c.sub.2 and c.sub.3 of
all of the matching branches are obtained.
[0074] In step 212, the matching scores, the difference signals and
the weight signals are synthesized as a fusion vector.
[0075] After the matching scores, the difference signals and the
weight signals are obtained, the matching scores, the difference
signals and the weight signals may be synthesized as the fusion
vector. For example, in a specific application scenario, the
synthesizing equation may be represented as:
x=[s.sub.1,s.sub.2,s.sub.3,d.sub.1,d.sub.2,d.sub.3,c.sub.1,c.sub.2,c.sub-
.3]
[0076] In step 213, the fusion vector is inputted into a
pre-trained SVM classifier to authenticate and determine whether
the chip image, the surface image and the live face image are
consistent with one another, to obtain an identity authentication
result.
[0077] After the synthesized fusion vector is obtained, the fusion
vector may be inputted into the pre-trained SVM classifier to
authenticate and determine whether the chip image, the surface
image and the live face image are consistent with one another, to
obtain an identity authentication result.
[0078] It should be illustrated that a SVM classifier may be
trained through the following steps A to D.
[0079] In step A, a chip image of an ID card, a surface image of
the ID card and a live face image are acquired as samples to
construct a ternary sample group in a training set. The ternary
sample group in the training set includes positive samples and
negative samples. For example, a ratio of the number of the
positive samples to the number of the negative samples may be
1:1.
[0080] In step B, a concatenated PCA convolution feature of each of
the images in the ternary sample group is extracted, to obtain
sample feature vectors.
[0081] In step C, information on the images in the ternary sample
group is fused based on the sample feature vectors according to the
score fusion strategy with the supervisory signal, to obtain a
sample fusion vector.
[0082] In step D, the sample fusion vector is inputted into the SVM
classifier for training the SVM classifier, to obtain the
pre-trained SVM classifier.
[0083] The above steps A, B, and C have the similar principle as
the above-described steps 201 to 212, and are not described
repeatedly herein anymore. It should be illustrated that the
ternary sample group in the training set includes positive samples
and negative samples. For example, a ratio of the number of the
positive samples to the number of the negative samples may be 1:1.
Three images in the triple sample group as positive samples
correspond to the same identity. That is, an authentication result
of the SVM classifier for the positive samples is that the
authentication is passed, and an output result of the SVM
classifier is 1. Three images in the triple sample group as
negative samples correspond to different identities, an
authentication result of the SVM classifier for the negative
samples is that the authentication is not passed, and an output
result of the SVM classifier is -1. For the negative samples, the
three images correspond to different identities as long as any one
of the three images is inconsistent with another image of the three
images. Therefore, there are many types of combination variation
for negative samples. In order to reduce redundancy of a sample
space, negative samples may be determined in the following ways. A
ternary sample group in which two of the chip image, the surface
image and the live face image correspond to the same identity, and
the other image than the two images correspond to a different
identity from the two images is selected as negative samples. It
should be understood that in a case where the three images
correspond to identities different from one another, the three
images may be determined as negative samples in any one "2+1" mode.
Therefore, the case where the three images correspond to identities
different from one another may be not taken into account in
constructing the sample space, thereby reducing redundancy of the
sample space, and improving learning efficiency of the SVM
classifier for the negative samples.
[0084] In addition, it should be understood that after the fusion
vector is inputted into the pre-trained SVM classifier, an
authentication result is provided by the SVM classifier. In a case
where an output value of the SVM classifier is 1, it indicates that
the three images correspond to the same identity and the identity
authentication is passed. In a case where the output value of the
SVM classifier is -1, it indicates the three images correspond to
different identities and the identity authentication is not
passed.
[0085] For ease of understanding, the method for offline identity
authentication according to an embodiment of the present disclosure
is described below through an application scenario with reference
to the embodiment described in FIG. 2.
[0086] In the application scenario, a flow of the method for
offline identity authentication is shown in FIG. 3, in which, the
SVM classifier is used as a classification decider, and a training
phase and an implementing (testing) phase are included.
[0087] In the training phase, samples are extracted from a training
image library to constitute appropriate ternary positive and
negative training samples. The training samples are inputted into
the SVM classifier through a concatenated PCA convolution feature
extracting module and a score fusing module with a supervisory
signal for training the SVM classifier. When constituting the
ternary positive and negative sample group, a combination of
positive samples is fixed, that is, a ternary group including three
images corresponding to the same identity is the combination of
positive samples. There are many combination variations for the
negative samples. In order to reduce redundancy of the sample
space, the "2+1" mode is used, that is, a ternary group in which
two of the three images correspond to the same identity and the
other image than the two images corresponds to a different identity
from the two images is selected as negative samples. The samples
are constituted as follows. A class label of the positive samples
is 1, and a class label of the negative samples is -1. The positive
samples are a ternary group of the same person in which the three
images correspond to the same identity feature and is represented
as [.sym., .sym., .circle-w/dot.]. The negative samples are a
ternary group in which at least one of the three images corresponds
to a different identity from other images of the three images, and
may be a combination such as [.beta., .sym., .circle-w/dot.],
[.sym., .circle-w/dot., .sym.)] and [.circle-w/dot., .sym., .sym.].
In a case where three images correspond to identities different
from one another, the three images may be determined as negative
samples in any one "2+1" mode. Therefore, the case may not be taken
into account in constructing the sample space, thereby reducing
redundancy of the sample space and improving learning efficiency of
the classifier for negative samples.
[0088] In the implementing (testing) phase, three face images are
acquired from the unified image acquisition module, and are input
into the pre-trained SVM classifier through the concatenated PCA
convolution feature extracting module and the score fusing module
with the supervisory signal, to identify and authenticate the three
face images. Since a feature extracting process and an information
fusing process in the training phase are the same as those in the
implementing (testing) phase, implementation of key technology of
the method in the implementing phase are described below.
[0089] In the implementing (testing) phase, the image acquisition
module outputs three face images, in which, a chip image may be
directly obtained by a card reading device, a surface image of an
ID card is obtained by a scanning device, and a live face image is
obtained by a camera device. A face region in the acquired image is
detected by a face detector. When acquiring the live face image,
the acquired face image is screened and filtered using a quality
evaluation algorithm, to ensure quality of the acquired image. In a
case where the acquired image does not meet a quality requirement,
an instruction may be automatically sent to instruct the holder to
re-acquire a face image. Preprocessing operations such as face
alignment and light correction are performed on the acquired image
to obtain a final output of the acquisition module.
[0090] In the feature extracting phase, convolution outputs of
multiple sub-layers are extracted in a pre-trained deep
convolutional neural network model based on the deep learning
algorithm, and PCA transform is performed layer by layer to obtain
the concatenated PCA convolution feature. The operating principle
is shown in FIG. 4. The pre-trained deep convolutional neural
network includes five convolution groups and two fully connected
layers. Each of the convolution groups includes two convolution
sub-layers and one pooling layer. When extracting the convolution
feature, outputs (pool4, pool5) of the pooling layers of the last
two convolution groups and an output (fc2) of the last fully
connected layer are extracted, and PCA transform is performed layer
by layer to reserve a principal component of the output of each
layer. The specific operation includes step (1) to (6) as
follows.
[0091] In step (1), an output pool4 of the polling layer of a
fourth convolution group is extracted, and all values of the output
are strung into a vector.
[0092] In step (2), PCA transform is performed on the vector pool4,
and the first n.sub.1 (for example n.sub.1=1024) principal
components are reserved to obtain PP1.
[0093] In step (3), an output pool5 of the pooling layer of a fifth
convolution group is extracted, and all values of the output are
strung into a vector, and PP1 is inserted into a header of the
vector.
[0094] In step (4), PCA transform is performed on the inserted
vector pool5, and the first n.sub.2 (for example, n.sub.2=512)
principal components are reserved to obtain PP2.
[0095] In step (5), an output fc2 of the second fully connected
layer is extracted, and PP2 is inserted into a header of the
vector.
[0096] In step (6), PCA transform is performed on the inserted
vector fc2, and the first n.sub.3 (for example, n.sub.3=256)
principal components are reserved to obtain PP3, which is a final
extracted concatenated PCA convolution feature.
[0097] In a score fusing phase, a score fusion strategy with a
supervisory signal is used. A basic principle of the score fusion
strategy is to construct two supervisory signals including a
difference signal and a weight signal based on a matching score,
and to join encode the ternary image group based on the two
supervisory signals in combination with the matching score. The
difference signal refers to a difference between a matching score
of each matching branch (a matching relationship between two images
represents a matching branch) and an empirical threshold of the
matching branch. The credibility of the matching score increases
with the increase in the difference. The weight signal is obtained
according to difference decision weights of the matching branches
by encoding based on the decision weight and the matching score. In
three-party authentication, environments in which the chip image is
acquired and the surface image of the ID card is acquired are
controllable, while image quality of the live face image subjects
to many uncontrollable factors such as an attitude, light and
shade. Therefore, in a process of joint comparison, the decision
weight of the matching branch between the chip image and the
surface image of the ID card may be large, while the decision
weights of the other two matching branches are small. The operating
principle of the score fusion strategy with the supervisory signal
is shown in FIG. 5, and a specific implementation includes step (1)
to (6) as follows.
[0098] In step (1), a concatenated PCA convolution feature of each
of the chip image, the surface image of the ID card and the live
face image is extracted based on the above process, and similarity
between each pair of the images is measured based on a cosine
similarity degree. The cosine similarity degree between two images
(I.sub.1, I.sub.2) is calculated according to the following
equation:
sim ( I 1 , I 2 ) = k = 1 n f 1 k f 2 k k = 1 n f 1 k 2 k = 1 n f 2
k 2 ##EQU00004##
[0099] where n denotes a dimension of the feature vector, f.sub.1k
denotes a k-th element of the feature vector of I.sub.1, and
f.sub.2k denotes a k-th element of the feature vector of I.sub.2.
According to the equation, a matching score between the chip image
and the surface image of the ID card is s.sub.1, and a matching
branch between the chip image and the surface image of the ID card
is denoted as a branch p1. A matching score between the chip image
and the live face image is s.sub.2, and a matching branch between
the chip image and the live face image is denoted as a branch p2. A
matching score between the surface image of the ID card and the
live face image is s.sub.3, and a matching branch between the
surface image of the ID card and the live face image is denoted as
a branch p3;
[0100] In step (2), an empirical threshold of each of the matching
branches in the training sample set is obtained using the 1:1
authentication algorithm, and the empirical threshold is calculated
according to the following equation:
arg max T i = 1 m .delta. { ( s i - T ) .times. y i > 0 } i = 1
m i ##EQU00005##
[0101] where m denotes the number of sample pairs, s.sub.i denotes
a similarity degree of an i-th sample pair, and y.sub.i is a class
label of the i-th sample pair. In a case where the sample pair
corresponds to the same person, the class label of the sample pair
is 1. In a case where the sample pair corresponds to different
persons, the class label of the sample pair is -1. .delta.{}
denotes an indicator function and is defined as follows:
.delta. { x } = { 1 , x is true 0 , x is false . ##EQU00006##
[0102] The empirical thresholds T.sub.1, T.sub.2 and T.sub.3 of the
three matching branches are obtained;
[0103] In step (3), the difference signal is calculated. The
empirical threshold of each of the matching branch is subtracted
from the matching score of the matching branch, to obtain
difference signals of the three matching branches, which are
d.sub.1=s.sub.1-T.sub.1, d.sub.2=s.sub.2-T.sub.2 and
d.sub.3=s.sub.3-T.sub.3.
[0104] In step 4, the weight signal is calculated. The weight
signal of each of the matching branches is represented with four
bits. The highest-order bit in the four bits is encoded based on
the difference signal and is determined according to
.delta.{(s-T)>0}. That is, the highest-order bit is encoded into
1 in a case of 5>T, and the highest-order bit is encoded into 0
in a case of s<T. The three lower-order bits in the four bits
are encoded based on the decision weight. The ratio of the decision
weights of the three matching branches is converted into an integer
ratio, and an integer in the integer ratio is normalized to a range
from 0 to 7. The decision weight of each of the matching branches
is encoded into three bits. For example, in a case where the ratio
of the decision weights of the branch p1, the branch p2 and the
branch p3 is 5:3:2, the decision weights are encoded into 101, 011,
and 010 respectively. In a case where the difference signal of the
branch p1 is positive, the highest-order bit of the four bits for
the branch p1 is 1, and the weight signal of the branch p1 is
encoded into binary codes 1101, and the binary codes are converted
into a decimal number 13. FIG. 6 shows a schematic encoding
diagram. Through the above operation, weight signals c.sub.1,
c.sub.2 and c.sub.3 of the matching branches are obtained.
[0105] In step 5, the matching scores, the difference signals and
the weight signals are synthesized as a final score fusion vector
x=[s.sub.1, s.sub.2, s.sub.3, d.sub.1, d.sub.2, d.sub.3, c.sub.1,
c.sub.2, c.sub.3] with a supervisory signal.
[0106] In step 6, in a decision phase, after the concatenated PCA
convolution features of the to-be-tested samples (the ternary group
including the chip image, the surface image of the ID card, and the
live face image which are acquired by the image acquisition module)
are extracted based on the above process, and the fusion vector
with a supervisory signal is generated based on the matching
scores, a determination result is provided by the pre-trained SVM
classifier automatically. In a case where the determination result
is 1, it indicates that the three images correspond to the same
identity and identity authentication is passed. In a case where the
determination result is -1, it indicates that the three images
correspond to different identities and identity authentication is
not passed.
[0107] In summary, with the offline identity authentication
algorithm in the present disclosure, whether a chip image of the ID
card, a surface image of the ID card and a live face image are
consistent with one another may be compared simultaneously offline,
thereby effectively implementing three-party identity
authentication based on the ID card, and especially effectively
solving the problem of a fake ID card which has a genuine chip and
a fake surface information. The three images are compared from an
overall decision. Whether the authentication is passed or not is
regarded as two classes to be decided, and a determination result
for the consistency is provided using a classification algorithm,
thereby avoiding a blurred decision caused by three times of
pairwise comparison as well as a decision rule in which a decision
priority level is set artificially. Specifically, the concatenated
PCA convolution feature having strong robustness and generalization
ability is provided based on the deep convolutional neural network,
to describe features of the images. A score fusion strategy with a
supervisory signal is provided in an information fusing phase,
thereby enriching pattern expression of the fusion vector in a
metric space, such that the classifier can fully learn a pattern
mapping relationship of the ternary image group and provides an
accurate determination result. As compared with the strategy in
which a decision is made simply based on a threshold and the
decision priority is set artificially, the score fusion strategy
with the supervisory signal is more intelligent and reliable.
[0108] The method for offline identity authentication is described
above. A device for offline identity authentication is described
below. Referring to FIG. 7, a device for offline identity
authentication according to an embodiment of the present disclosure
includes a multivariate image acquiring module 701, a convolution
feature extracting module 702, a score fusing module 703 and an
authenticating and determining module 704
[0109] The multivariate image acquiring module 701 is configured to
acquire two or more images for identity authentication, to
constitute a to-be-authenticated multivariate image group.
[0110] The convolution feature extracting module 702 is configured
to extract a concatenated PCA convolution feature of each of the
images in the to-be-authenticated multivariate image group, to
obtain feature vectors.
[0111] The score fusing module 703 is configured to fuse
information on the images in the to-be-authenticated multivariate
image group based on the feature vectors according to a score
fusion strategy with a supervisory signal, to obtain a fusion
vector.
[0112] The authenticating and determining module 704 is configured
to input the fusion vector into a pre-trained SVM classifier to
authenticate and determine whether the chip image, the surface
image and the live face image are consistent with one another, to
obtain an identity authentication result.
[0113] In this embodiment, the multivariate image acquiring module
701 acquires two or more images for identity authentication to
constitute a to-be-authenticated multivariate image group. The
convolution feature extracting module 702 extracts a concatenated
PCA convolution feature of each of the images in the
to-be-authenticated multivariate image group to obtain feature
vectors. The score fusing module 703 fuses information on the
images in the to-be-authenticated multivariate image group based on
the feature vectors according to a score fusion strategy with a
supervisory signal, to obtain a fusion vector. The authenticating
and determining module 704 inputs the fusion vector into a
pre-trained SVM classifier to authenticate and determine whether
the images in the to-be-authenticated multivariate image group are
consistent with one another, to obtain an identity authentication
result. In this embodiment, images for identity authentication are
compared offline from an overall decision, and an authentication
result is provided by a single decision of the pre-trained SVM
classifier, thereby reducing burden on the checker and improving
authentication efficiency without relying on face database of the
Ministry of Public Security.
[0114] For ease of understanding, a device for offline identity
authentication according to an embodiment of the present disclosure
is described in detail below. Referring to FIG. 8, a device for
offline identity authentication according to another embodiment of
the present disclosure includes a multivariate image acquiring
module 801, a convolution feature extracting module 802, a score
fusing module 803 and an authenticating and determining module
804.
[0115] The multivariate image acquiring module 801 is configured to
acquire two or more images for identity authentication to
constitute a to-be-authenticated multivariate image group.
[0116] The convolution feature extracting module 802 is configured
to extract a concatenated PCA convolution feature of each of the
images in the to-be-authenticated multivariate image group to
obtain feature vectors.
[0117] The score fusing module 803 is configured to fuse
information on the images in the to-be-authenticated multivariate
image group based on the feature vectors according to a score
fusion strategy with a supervisory signal, to obtain a fusion
vector.
[0118] The authenticating and determining module 804 is configured
to input the fusion vector into a pre-trained SVM classifier to
authenticate and determine whether the images in the
to-be-authenticated multivariate image group are consistent with
one another, to obtain an identity authentication result.
[0119] In this embodiment, the multivariate image group may be a
ternary image group, and include a chip image of an ID card, a
surface image of the ID card and a live face image.
[0120] In this embodiment, a SVM classifier may be trained by a
ternary sample acquiring module 805, a sample convolution feature
extracting module 806, a sample score fusing module 807 and a
classifier training module 808.
[0121] The ternary sample acquiring module 805 is configured to
acquire a chip image of an ID card, a surface image of the ID card
and a live face image as samples to constitute a ternary sample
group in a training set. The ternary sample group in the training
set includes positive samples and negative samples. For example, a
ratio of the number of the positive samples to the number of the
negative samples may be 1:1.
[0122] The sample convolution feature extracting module 806 is
configured to extract a concatenated PCA convolution feature of
each of the images in the ternary sample group to obtain sample
feature vectors.
[0123] The sample score fusing module 807 is configured to fuse
information on the images in the ternary sample group based on the
sample feature vectors according to a score fusion strategy with a
supervisory signal, to obtain a sample fusion vector.
[0124] The classifier training module 808 is configured to input
the sample fusion vector to the SVM classifier for training the SVM
classifier, to obtain the pre-trained SVM classifier.
[0125] In this embodiment, the score fusing module 803 includes a
matching score calculating unit 8031, a difference signal
calculating unit 8032, a weight signal calculating unit 8033 and a
fusion vector synthesizing unit 8034.
[0126] The matching score calculating unit 8031 is configured to
calculate a cosine similarity degree between each pair of feature
vectors of the three feature vectors corresponding to the
to-be-authenticated ternary image group, as three matching
scores.
[0127] The difference signal calculating unit 8032 is configured to
calculate a difference between each of the matching scores and a
preset empirical threshold corresponding to the matching score, as
difference signals.
[0128] The weight signal calculating unit 8033 is configured to
encode based on each of preset decision weights and the difference
signal corresponding to the preset decision weight, to obtain
weight signals. The preset decision weights have a one-to-one
correspondence with three pairs of images in the
to-be-authenticated ternary image group.
[0129] The fusion vector synthesizing unit 8034 is configured to
synthesize the matching scores, the difference signals and the
weight signals, as a fusion vector.
[0130] It may be clearly understood by those skilled in the art
that, for convenience and ease of description, operating processes
of the system, the device and the unit described above may refer to
the corresponding processes in the above method embodiments, which
are not described repeatedly here anymore.
[0131] In several embodiments according to the present disclosure,
it should be understood that the disclosed system, device and
method can be implemented in other ways. The device embodiments
described above are merely schematic. For example, the division of
the units is merely a logic functional division, and there may be
other divisions in practice. For example, multiple units or
components may be combined, or may be integrated into another
system, or some features may be ignored or not be executed. In
addition, coupling, direct coupling or communication connection
between components shown or discussed may be indirect coupling or
communication connection via some interfaces, devices or units,
which may be electrical, mechanical, or in other form.
[0132] The units illustrated as separate components may be or may
not be separated physically, and the component displayed as a unit
may be or may not be a physical unit. That is, the components may
be located at the same place, or may be distributed on multiple
network units. Some or all of the units may be selected as required
to implement the objective of the solution of the embodiments.
[0133] In addition, all function units according to the embodiments
of the present disclosure may be integrated into one processing
unit, or may be each a separate unit physically, or two or more
units are integrated into one unit. The integrated unit described
above may be realized with hardware, or may be realized by a
software function unit.
[0134] The integrated unit may be stored in a computer readable
storage medium if the integrated unit is implemented in the form of
a software function unit and is sold or used as a separate product.
Base on such understanding, an essential part of the technical
solution of the present disclosure, i.e., the part of the technical
solution of the present disclosure that contributes to the
conventional technology, or all or a part of the technical solution
may be embodied in the form of a computer software product. The
computer software product is stored in a storage medium, and
includes several instructions to instructing a computer device
(which may be a personal computer, a server, a network device or
the like) to implement all or a part of steps of the method
according to the embodiments of the present disclosure. The storage
medium described above includes various mediums which can store
program codes such as a USB disk, a mobile hard disk, a read-only
memory (ROM), a random access memory (RAM), a magnetic disk and an
optical disc.
[0135] In summary, the above embodiments are only described for
illustrating the technical solutions of the present disclosure, and
not for limiting the technical solutions. Although the present
disclosure is illustrated in detail by referring to the
aforementioned embodiments, it should be understood by those
skilled in the art that modifications can be still made to the
technical solutions recited in the aforementioned embodiments, or
equivalent substitution can be made to a part of technical features
of the technical solutions. The modification and equivalent
substitution cannot make essence of the technical solutions depart
from the spirit and a scope of the technical solutions according to
the embodiments of the present disclosure.
* * * * *